By ByteBot
On October 15, Microsoft AI completed pre-training of MAI-Base-2 — a 2.8-trillion-parameter mixture-of-experts model trained on 16,000 GB200 GPUs over 38 days. The company claims it outperforms every open-weight model on its internal benchmarks and runs 4 to 12 times more FLOPs-efficiently than DeepSeek-V4-Pro-Base, the strongest open base model available. Here’s what Microsoft didn’t put in the headline: there are no weights to download, no API to call, no release date, and no published price. For most developers, Microsoft MAI-Base-2 is a benchmark claim with nothing callable behind it.
What Microsoft MAI-Base-2 Actually Is
MAI-Base-2 uses a sparse mixture-of-experts architecture: 2.8 trillion total parameters, but only 80 billion activate per token. The model routes each token through 7 of 512 expert networks plus one shared expert — that’s how it achieves massive theoretical capacity while keeping compute costs closer to an 80B dense model. Microsoft trained it on 50 trillion tokens (36% more than MAI-Base-1), incorporating 4.5 trillion vision tokens representing 18% of the mix, making MAI-Base-2 natively multimodal without a separate image encoder.
The training run included several engineering improvements worth noting. Microsoft replaced AdamW with the Muon optimizer, which delivered more stable training dynamics at just 2.1% of iteration time. A “dropless” MoE routing approach handles token overflow across multiple rounds instead of dropping tokens — fixing a common failure mode at this scale. The 89% goodput figure (time spent on actual training versus hardware restarts) is strong for a run of this scale on 16,000 GB200 GPUs over 38 days. According to Microsoft’s official announcement, MAI-Base-3 preparations are already underway.
The Access Problem
MAI-Base-2 is a base checkpoint — not a finished model. Base models cannot follow instructions, use tools, or respond usefully to queries. They require post-training: instruction tuning, RLHF, safety filtering. That work is currently underway at Microsoft, with no announced completion date. Announcing a base checkpoint as a model release is roughly equivalent to announcing you’ve built a car engine. Useful when it’s eventually attached to something, but not something you can drive home today.
Even after post-training completes, access won’t be straightforward. Microsoft plans to license MAI-Base-2 weights exclusively to “trusted partners” through its Foundry enterprise program. No public weights, no Hugging Face repository, no open API. As one developer analysis put it: “No repository, no checkpoint, no quantisation, no download of any kind.” Microsoft’s own blog confirms the model is in post-training with no timeline given.
Related: Meta Muse Spark Open: 816B Agent Model, One License Gotcha
Open in Name Only
The benchmarking framing deserves scrutiny. Microsoft evaluated MAI-Base-2 against genuinely open-weight models — DeepSeek-V4-Pro-Base, Kimi K3, Qwen3.8-2.4T-A95B, MiMo-V2.6-Pro-RL — and claimed superiority on held-out bits-per-byte evaluations. However, those comparisons are Microsoft’s internal benchmarks, not reproducible public suites. Some comparisons pit MAI-Base-2’s base checkpoint against rivals’ post-trained checkpoints, which is not a fair comparison. Furthermore, the model that “beats” open-weight competitors won’t itself be available as open weights.
Constellation Research described this as an emerging “open-ish” category — not fully open, not frontier-grade, but commercially accessible through enterprise agreements. That framing is honest. Microsoft’s strategy is explicit: build MAI-Base-2 into its own products first, then make it available to Foundry partners. The open-source community that built DeepSeek, Llama, and Qwen gets nothing from this announcement.
What You Can Actually Use Today
On the same day Microsoft announced MAI-Base-2, Nvidia shipped Nemotron 3.5 Super VL: a 120B MoE model (12B active parameters) with a 1 million-token context window, native support for text, images, documents, and video — and actual public weights on Hugging Face. It’s available through Nvidia’s NIM microservices, OpenRouter, ModelScope, and build.nvidia.com. Meanwhile, Nemotron 3.5 Super VL weights are downloadable today. Red Hat AI released optimized FP8 checkpoints on launch day.
The contrast between the two October 15 announcements is instructive. MAI-Base-2 is technically more ambitious: 2.8T parameters versus 120B, a longer training run, more novel architectural choices. Nevertheless, Nemotron 3.5 Super VL is usable. For most developers deciding what to integrate this week, that gap matters more than the parameter count.
Key Takeaways
- MAI-Base-2 completed pre-training on October 15 — 2.8T parameters, 80B active, trained on 16,000 GB200 GPUs. Technically a significant engineering achievement.
- Nothing is callable yet. The model is in post-training with no release date, no API, and no public weights. Foundry enterprise licensing is the only access path — for “trusted partners” only.
- Microsoft’s efficiency claims (4-12x over DeepSeek) are based on internal benchmarks, not reproducible public evaluations. Treat them as directional until independent verification exists.
- If you need a large open multimodal model today, Nvidia’s Nemotron 3.5 Super VL (120B, 1M context, public weights on Hugging Face) launched the same day and is available now.
- Watch Microsoft Foundry for post-training completion news. If you’re on an Azure enterprise agreement, that’s where MAI-Base-2 access will materialize first.
